Google DeepMind has unveiled Gemini 4 Argon, a model it describes as a step change in physical understanding, code generation, and visual-semantic reasoning. The company says the model is designed to unlock new workflows across the robotics stack and accelerate progress toward physical AGI.
The model's capability is measured on the Embodied Reasoning (ERQA) benchmark, which tests how well a model understands and solves physical, spatial problems in 3D environments. Using pure reasoning, Argon scores 87%, outperforming every other frontier model. DeepMind positions it as a general-purpose physical reasoning model that can be applied across the entire robotics lifecycle.
DeepMind's examples fall into four areas: robot control, Real2Sim, physical understanding, and research and self-improvement. In one demonstration, Argon controls a Boston Dynamics Spot quadruped with a six-degree-of-freedom arm to learn a new skill. It builds an articulated MuJoCo digital twin of a foot-pedal trash bin, practices its moves in simulation, then steps on the physical pedal and drops soda cans into the bin.
In a zero-shot control example, Argon directs a simple robotic arm to set a fallen mug upright and place a crumpled napkin inside. It reasons with a visual coordinate harness and action primitives over several turns to complete the task. The examples show how the model can move from simulation to real hardware with minimal task-specific training.
Source: Google DeepMind Blog · Summarized by HeadlinesBriefing